{
  "id": 243356,
  "title": "Learnings From the Competition ,  39th Place Solution",
  "url": "/competitions/birdclef-2021/writeups/titans-learnings-from-the-competition-39th-place-s",
  "author_name": "",
  "post_date": "2021-06-02T20:16:09.277Z",
  "votes": 11,
  "comment_count": 7,
  "views": 0,
  "content": "<p>Hi all,<br>\nI really wish I could have started this competition sooner but I did the best I could in 10 days . It never feels good to go down in private even if its 1 place , but what really was more disappointing was that we failed to pick our best submission which was giving <b> 0.66 </b> on private. It was an absolute grind for us , tons of hard work</p>\n<p>Our bird didn't fly but it was a great experience . <b> I would like to thank <a href=\"https://www.kaggle.com/nischaydnk\" target=\"_blank\">@nischaydnk</a> for agreeing to join this competition this late and putting in immense work , an absolute gem to have as a teammate </b> and my other teammate <a href=\"https://www.kaggle.com/parthdhameliya77\" target=\"_blank\">@parthdhameliya77</a> , <a href=\"https://www.kaggle.com/vatsal\" target=\"_blank\">@vatsal</a> and <a href=\"https://www.kaggle.com/drthrevan\" target=\"_blank\">@drthrevan</a> for accepting us to the team</p>\n<p>It was my first serious experience with audio data and <a href=\"https://www.youtube.com/watch?v=iCwMQJnKk2c&amp;list=PL-wATfeyAMNqIee7cH3q1bh4QJFAaeNv0\" target=\"_blank\">this</a> youtube playlist helped me get up to speed and really get the hang of audio data. Thanks to kaggle community and winners of the previous Birdcall competition sharing their solutions , it was easy for us to get the baselines up and ready</p>\n<h1>Things that worked</h1>\n<ul>\n<li>Using Delta for stacking instead of stacking the same spectograms 3 times ( Idea taken from Birdcall's 4th place solution )</li>\n<li>RMS for cropping audio instead of random cropping  ( Idea taken from Birdcall's 4th place solution)</li>\n<li>Resnext and Nfnet series of models</li>\n<li>Post Processing </li>\n</ul>\n<h1>Things That didn't work</h1>\n<ul>\n<li><p>I tried to mimic last year's 1st place solution SED model <a href=\"https://www.kaggle.com/tanulsingh077/sed-model-tanul-s-version\" target=\"_blank\">here</a> but the results weren't great because one epoch was an hour long and it took atleast 35 epochs to converge , we were outdone by the limited hardware power and time here . But we learned the in and out absolute working of the model</p></li>\n<li><p>CPMP's sampling method : Instead of taking random crops , crop from the begining and the end</p></li>\n<li><p>Efficientnet Family of models , I tried to mimic last year's fourth place training strategy but it didn't work out</p></li>\n<li><p>Longer time durations (10 secs , 15 secs melspecs) , different temporal dimensions (changing hop size and frame size)</p></li>\n<li><p>Vision transformer models , we used code provided by the organizers to generate a square mel specs but that mess up the temporal dimension I guess and hence the results were not that good</p></li>\n</ul>\n<h1>Things that we missed</h1>\n<ul>\n<li><p>We used already saved mel spectograms to train our models as we were limited by hardware and time and hence we were not able to use any of the audio level transforms and now I feel we missed the most there .<b> The key was to make the model robust to noise and overfit less </b></p></li>\n<li><p>Failing to use the meta-information given to us in terms of location and time</p></li>\n</ul>\n<h1>Our Final Solution Summary</h1>\n<p>Our final model consisted of Nfnet_l0 (5folds) and 2 resnext50 (5folds) (with and without delta stacking)trained on 7 secs mel specs ,sigmoid level ensemble with a post-processing layer . The post-processing was something which gave us the biggest boost , it was <a href=\"https://www.kaggle.com/nischaydnk\" target=\"_blank\">@nischaydnk</a> great idea , so I will let him explain it in the comments.</p>\n<p>I take away a lot of learnings from this competition and I will be waiting eagerly for the next audio competition </p>",
  "messages": [
    {
      "id": "1332580",
      "postDate": "06/02/2021 07:43:14",
      "content": "<p>Hi all,<br>\nI really wish I could have started this competition sooner but I did the best I could in 10 days . It never feels good to go down in private even if its 1 place , but what really was more disappointing was that we failed to pick our best submission which was giving <b> 0.66 </b> on private. It was an absolute grind for us , tons of hard work</p>\n<p>Our bird didn't fly but it was a great experience . <b> I would like to thank <a href=\"https://www.kaggle.com/nischaydnk\" target=\"_blank\">@nischaydnk</a> for agreeing to join this competition this late and putting in immense work , an absolute gem to have as a teammate </b> and my other teammate <a href=\"https://www.kaggle.com/parthdhameliya77\" target=\"_blank\">@parthdhameliya77</a> , <a href=\"https://www.kaggle.com/vatsal\" target=\"_blank\">@vatsal</a> and <a href=\"https://www.kaggle.com/drthrevan\" target=\"_blank\">@drthrevan</a> for accepting us to the team</p>\n<p>It was my first serious experience with audio data and <a href=\"https://www.youtube.com/watch?v=iCwMQJnKk2c&amp;list=PL-wATfeyAMNqIee7cH3q1bh4QJFAaeNv0\" target=\"_blank\">this</a> youtube playlist helped me get up to speed and really get the hang of audio data. Thanks to kaggle community and winners of the previous Birdcall competition sharing their solutions , it was easy for us to get the baselines up and ready</p>\n<h1>Things that worked</h1>\n<ul>\n<li>Using Delta for stacking instead of stacking the same spectograms 3 times ( Idea taken from Birdcall's 4th place solution )</li>\n<li>RMS for cropping audio instead of random cropping  ( Idea taken from Birdcall's 4th place solution)</li>\n<li>Resnext and Nfnet series of models</li>\n<li>Post Processing </li>\n</ul>\n<h1>Things That didn't work</h1>\n<ul>\n<li><p>I tried to mimic last year's 1st place solution SED model <a href=\"https://www.kaggle.com/tanulsingh077/sed-model-tanul-s-version\" target=\"_blank\">here</a> but the results weren't great because one epoch was an hour long and it took atleast 35 epochs to converge , we were outdone by the limited hardware power and time here . But we learned the in and out absolute working of the model</p></li>\n<li><p>CPMP's sampling method : Instead of taking random crops , crop from the begining and the end</p></li>\n<li><p>Efficientnet Family of models , I tried to mimic last year's fourth place training strategy but it didn't work out</p></li>\n<li><p>Longer time durations (10 secs , 15 secs melspecs) , different temporal dimensions (changing hop size and frame size)</p></li>\n<li><p>Vision transformer models , we used code provided by the organizers to generate a square mel specs but that mess up the temporal dimension I guess and hence the results were not that good</p></li>\n</ul>\n<h1>Things that we missed</h1>\n<ul>\n<li><p>We used already saved mel spectograms to train our models as we were limited by hardware and time and hence we were not able to use any of the audio level transforms and now I feel we missed the most there .<b> The key was to make the model robust to noise and overfit less </b></p></li>\n<li><p>Failing to use the meta-information given to us in terms of location and time</p></li>\n</ul>\n<h1>Our Final Solution Summary</h1>\n<p>Our final model consisted of Nfnet_l0 (5folds) and 2 resnext50 (5folds) (with and without delta stacking)trained on 7 secs mel specs ,sigmoid level ensemble with a post-processing layer . The post-processing was something which gave us the biggest boost , it was <a href=\"https://www.kaggle.com/nischaydnk\" target=\"_blank\">@nischaydnk</a> great idea , so I will let him explain it in the comments.</p>\n<p>I take away a lot of learnings from this competition and I will be waiting eagerly for the next audio competition </p>",
      "rawMarkdown": "Hi all,\nI really wish I could have started this competition sooner but I did the best I could in 10 days . It never feels good to go down in private even if its 1 place , but what really was more disappointing was that we failed to pick our best submission which was giving <b> 0.66 </b> on private. It was an absolute grind for us , tons of hard work\n\nOur bird didn't fly but it was a great experience . <b> I would like to thank @nischaydnk for agreeing to join this competition this late and putting in immense work , an absolute gem to have as a teammate </b> and my other teammate @parthdhameliya77 , @vatsal and @drthrevan for accepting us to the team\n\nIt was my first serious experience with audio data and [this](https://www.youtube.com/watch?v=iCwMQJnKk2c&list=PL-wATfeyAMNqIee7cH3q1bh4QJFAaeNv0) youtube playlist helped me get up to speed and really get the hang of audio data. Thanks to kaggle community and winners of the previous Birdcall competition sharing their solutions , it was easy for us to get the baselines up and ready\n\n# Things that worked\n\n*  Using Delta for stacking instead of stacking the same spectograms 3 times ( Idea taken from Birdcall's 4th place solution )\n*  RMS for cropping audio instead of random cropping  ( Idea taken from Birdcall's 4th place solution)\n*  Resnext and Nfnet series of models\n*  Post Processing \n\n# Things That didn't work\n\n* I tried to mimic last year's 1st place solution SED model [here](https://www.kaggle.com/tanulsingh077/sed-model-tanul-s-version) but the results weren't great because one epoch was an hour long and it took atleast 35 epochs to converge , we were outdone by the limited hardware power and time here . But we learned the in and out absolute working of the model\n\n* CPMP's sampling method : Instead of taking random crops , crop from the begining and the end\n\n* Efficientnet Family of models , I tried to mimic last year's fourth place training strategy but it didn't work out\n\n* Longer time durations (10 secs , 15 secs melspecs) , different temporal dimensions (changing hop size and frame size)\n\n* Vision transformer models , we used code provided by the organizers to generate a square mel specs but that mess up the temporal dimension I guess and hence the results were not that good\n\n# Things that we missed \n\n* We used already saved mel spectograms to train our models as we were limited by hardware and time and hence we were not able to use any of the audio level transforms and now I feel we missed the most there .<b> The key was to make the model robust to noise and overfit less </b>\n\n* Failing to use the meta-information given to us in terms of location and time\n\n# Our Final Solution Summary\n\nOur final model consisted of Nfnet_l0 (5folds) and 2 resnext50 (5folds) (with and without delta stacking)trained on 7 secs mel specs ,sigmoid level ensemble with a post-processing layer . The post-processing was something which gave us the biggest boost , it was @nischaydnk great idea , so I will let him explain it in the comments.\n\n\nI take away a lot of learnings from this competition and I will be waiting eagerly for the next audio competition",
      "votes": null
    },
    {
      "id": "1332591",
      "postDate": "06/02/2021 07:49:28",
      "content": "<p>Thanks <a href=\"https://www.kaggle.com/tanulsingh077\" target=\"_blank\">@tanulsingh077</a> for sharing this!!<br>\nAlso the reference in youtube is very helpful… the way I struggled in the competition even after looking at previous competition solutions made me realise I need to go back to the basics and have a good understanding of audio ML. This will be a useful reference! <br>\ncongrats to you and the team on silver position!</p>",
      "rawMarkdown": "Thanks @tanulsingh077 for sharing this!!\nAlso the reference in youtube is very helpful... the way I struggled in the competition even after looking at previous competition solutions made me realise I need to go back to the basics and have a good understanding of audio ML. This will be a useful reference! \ncongrats to you and the team on silver position!",
      "votes": null
    },
    {
      "id": "1335401",
      "postDate": "06/04/2021 07:32:53",
      "content": "<p>Congratulation <a href=\"https://www.kaggle.com/tanulsingh077\" target=\"_blank\">@tanulsingh077</a> for your 39th place and thanks for the detailed writing.</p>\n<p>The idea of using deltas for stacking seems quite interesting. In terms of CV f1 points, how much did it helps? Also, did the two backbones models Resnext and Nfnet performed about the same, or one was better?</p>",
      "rawMarkdown": "Congratulation @tanulsingh077 for your 39th place and thanks for the detailed writing.\n\nThe idea of using deltas for stacking seems quite interesting. In terms of CV f1 points, how much did it helps? Also, did the two backbones models Resnext and Nfnet performed about the same, or one was better?",
      "votes": null
    },
    {
      "id": "1335956",
      "postDate": "06/04/2021 14:44:04",
      "content": "<p>Thanks for sharing.</p>\n<blockquote>\n  <p>what really was more disappointing was that we failed to pick our best submission which was giving 0.66 on private</p>\n</blockquote>\n<p>How would you have known this?  Have you learned something now that would have let you pick this submission?</p>\n<p>It is interesting to see that 3 of the things that did not work work you did marvel for me.  Devil is in details probably.</p>",
      "rawMarkdown": "Thanks for sharing.\n\n> what really was more disappointing was that we failed to pick our best submission which was giving 0.66 on private\n\nHow would you have known this?  Have you learned something now that would have let you pick this submission?\n\nIt is interesting to see that 3 of the things that did not work work you did marvel for me.  Devil is in details probably.",
      "votes": null
    },
    {
      "id": "1336011",
      "postDate": "06/04/2021 15:27:53",
      "content": "<blockquote>\n  <p>How would you have known this? Have you learned something now that would have let you pick this submission?</p>\n</blockquote>\n<p>Not before reading Psi's solution , our CV was same as everyone else's we used the train soundscapes 20 files to evaluate and chose the best threshold , but since we were training using 7 second clips and were also selecting the clips randomly , moreover our cv on train soundscapes correlated well with the lb , that made us careless and we didn't think much</p>\n<blockquote>\n  <p>It is interesting to see that 3 of the things that did not work work you did marvel for me. Devil is in details probably.</p>\n</blockquote>\n<p>You are right devil is in the details , we might have made it work if only we had time , but I learned a lot and I am fortunate that I participated ,maybe I can do well in next year's competition</p>",
      "rawMarkdown": "> How would you have known this? Have you learned something now that would have let you pick this submission?\n\nNot before reading Psi's solution , our CV was same as everyone else's we used the train soundscapes 20 files to evaluate and chose the best threshold , but since we were training using 7 second clips and were also selecting the clips randomly , moreover our cv on train soundscapes correlated well with the lb , that made us careless and we didn't think much\n\n> It is interesting to see that 3 of the things that did not work work you did marvel for me. Devil is in details probably.\n\nYou are right devil is in the details , we might have made it work if only we had time , but I learned a lot and I am fortunate that I participated ,maybe I can do well in next year's competition",
      "votes": null
    },
    {
      "id": "1336013",
      "postDate": "06/04/2021 15:29:29",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/jonathanbesomi\" target=\"_blank\">@jonathanbesomi</a> Using Delta our resnext cv improved from 0.712 --&gt; 0.722 and our nfnet's cv increased from 0.728 ---&gt; 0.730 </p>\n<p>Nfnet was better than resnext , but their ensemble did not work for some reason and it gave bad cv as well as lb</p>",
      "rawMarkdown": "Hi @jonathanbesomi Using Delta our resnext cv improved from 0.712 --> 0.722 and our nfnet's cv increased from 0.728 ---> 0.730 \n\nNfnet was better than resnext , but their ensemble did not work for some reason and it gave bad cv as well as lb",
      "votes": null
    },
    {
      "id": "1336014",
      "postDate": "06/04/2021 15:29:55",
      "content": "<p>Thanks <a href=\"https://www.kaggle.com/kamal\" target=\"_blank\">@kamal</a> basics are always the key to win kaggle competition</p>",
      "rawMarkdown": "Thanks @kamal basics are always the key to win kaggle competition",
      "votes": null
    },
    {
      "id": "1336432",
      "postDate": "06/04/2021 22:38:02",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/tanulsingh077\" target=\"_blank\">@tanulsingh077</a>, thanks for your reply. Wow, 0.01 improvement is really a lot! I've been trying training different models (similar to kkiller notebook) but had hard time to beat the 0.7 CV score. Would you mind make your code public? I would like to train one of you models to understand where I was missing out. Also, I'm pretty sure other Kagglers will find this useful as well. Thanks!</p>",
      "rawMarkdown": "Hi @tanulsingh077, thanks for your reply. Wow, 0.01 improvement is really a lot! I've been trying training different models (similar to kkiller notebook) but had hard time to beat the 0.7 CV score. Would you mind make your code public? I would like to train one of you models to understand where I was missing out. Also, I'm pretty sure other Kagglers will find this useful as well. Thanks!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1332591,
      "author_name": "kmldas",
      "author_url": "",
      "post_date": "06/02/2021 07:49:28",
      "content": "<p>Thanks <a href=\"https://www.kaggle.com/tanulsingh077\" target=\"_blank\">@tanulsingh077</a> for sharing this!!<br>\nAlso the reference in youtube is very helpful… the way I struggled in the competition even after looking at previous competition solutions made me realise I need to go back to the basics and have a good understanding of audio ML. This will be a useful reference! <br>\ncongrats to you and the team on silver position!</p>",
      "votes": null,
      "replies": [
        {
          "id": 1336014,
          "author_name": "tanulsingh077",
          "author_url": "",
          "post_date": "06/04/2021 15:29:55",
          "content": "<p>Thanks <a href=\"https://www.kaggle.com/kamal\" target=\"_blank\">@kamal</a> basics are always the key to win kaggle competition</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1335401,
      "author_name": "jonathanbesomi",
      "author_url": "",
      "post_date": "06/04/2021 07:32:53",
      "content": "<p>Congratulation <a href=\"https://www.kaggle.com/tanulsingh077\" target=\"_blank\">@tanulsingh077</a> for your 39th place and thanks for the detailed writing.</p>\n<p>The idea of using deltas for stacking seems quite interesting. In terms of CV f1 points, how much did it helps? Also, did the two backbones models Resnext and Nfnet performed about the same, or one was better?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1336013,
          "author_name": "tanulsingh077",
          "author_url": "",
          "post_date": "06/04/2021 15:29:29",
          "content": "<p>Hi <a href=\"https://www.kaggle.com/jonathanbesomi\" target=\"_blank\">@jonathanbesomi</a> Using Delta our resnext cv improved from 0.712 --&gt; 0.722 and our nfnet's cv increased from 0.728 ---&gt; 0.730 </p>\n<p>Nfnet was better than resnext , but their ensemble did not work for some reason and it gave bad cv as well as lb</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1336432,
          "author_name": "jonathanbesomi",
          "author_url": "",
          "post_date": "06/04/2021 22:38:02",
          "content": "<p>Hi <a href=\"https://www.kaggle.com/tanulsingh077\" target=\"_blank\">@tanulsingh077</a>, thanks for your reply. Wow, 0.01 improvement is really a lot! I've been trying training different models (similar to kkiller notebook) but had hard time to beat the 0.7 CV score. Would you mind make your code public? I would like to train one of you models to understand where I was missing out. Also, I'm pretty sure other Kagglers will find this useful as well. Thanks!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1335956,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "06/04/2021 14:44:04",
      "content": "<p>Thanks for sharing.</p>\n<blockquote>\n  <p>what really was more disappointing was that we failed to pick our best submission which was giving 0.66 on private</p>\n</blockquote>\n<p>How would you have known this?  Have you learned something now that would have let you pick this submission?</p>\n<p>It is interesting to see that 3 of the things that did not work work you did marvel for me.  Devil is in details probably.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1336011,
          "author_name": "tanulsingh077",
          "author_url": "",
          "post_date": "06/04/2021 15:27:53",
          "content": "<blockquote>\n  <p>How would you have known this? Have you learned something now that would have let you pick this submission?</p>\n</blockquote>\n<p>Not before reading Psi's solution , our CV was same as everyone else's we used the train soundscapes 20 files to evaluate and chose the best threshold , but since we were training using 7 second clips and were also selecting the clips randomly , moreover our cv on train soundscapes correlated well with the lb , that made us careless and we didn't think much</p>\n<blockquote>\n  <p>It is interesting to see that 3 of the things that did not work work you did marvel for me. Devil is in details probably.</p>\n</blockquote>\n<p>You are right devil is in the details , we might have made it work if only we had time , but I learned a lot and I am fortunate that I participated ,maybe I can do well in next year's competition</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1332580": "Hi all,\nI really wish I could have started this competition sooner but I did the best I could in 10 days . It never feels good to go down in private even if its 1 place , but what really was more disappointing was that we failed to pick our best submission which was giving <b> 0.66 </b> on private. It was an absolute grind for us , tons of hard work\n\nOur bird didn't fly but it was a great experience . <b> I would like to thank @nischaydnk for agreeing to join this competition this late and putting in immense work , an absolute gem to have as a teammate </b> and my other teammate @parthdhameliya77 , @vatsal and @drthrevan for accepting us to the team\n\nIt was my first serious experience with audio data and [this](https://www.youtube.com/watch?v=iCwMQJnKk2c&list=PL-wATfeyAMNqIee7cH3q1bh4QJFAaeNv0) youtube playlist helped me get up to speed and really get the hang of audio data. Thanks to kaggle community and winners of the previous Birdcall competition sharing their solutions , it was easy for us to get the baselines up and ready\n\n# Things that worked\n\n*  Using Delta for stacking instead of stacking the same spectograms 3 times ( Idea taken from Birdcall's 4th place solution )\n*  RMS for cropping audio instead of random cropping  ( Idea taken from Birdcall's 4th place solution)\n*  Resnext and Nfnet series of models\n*  Post Processing \n\n# Things That didn't work\n\n* I tried to mimic last year's 1st place solution SED model [here](https://www.kaggle.com/tanulsingh077/sed-model-tanul-s-version) but the results weren't great because one epoch was an hour long and it took atleast 35 epochs to converge , we were outdone by the limited hardware power and time here . But we learned the in and out absolute working of the model\n\n* CPMP's sampling method : Instead of taking random crops , crop from the begining and the end\n\n* Efficientnet Family of models , I tried to mimic last year's fourth place training strategy but it didn't work out\n\n* Longer time durations (10 secs , 15 secs melspecs) , different temporal dimensions (changing hop size and frame size)\n\n* Vision transformer models , we used code provided by the organizers to generate a square mel specs but that mess up the temporal dimension I guess and hence the results were not that good\n\n# Things that we missed \n\n* We used already saved mel spectograms to train our models as we were limited by hardware and time and hence we were not able to use any of the audio level transforms and now I feel we missed the most there .<b> The key was to make the model robust to noise and overfit less </b>\n\n* Failing to use the meta-information given to us in terms of location and time\n\n# Our Final Solution Summary\n\nOur final model consisted of Nfnet_l0 (5folds) and 2 resnext50 (5folds) (with and without delta stacking)trained on 7 secs mel specs ,sigmoid level ensemble with a post-processing layer . The post-processing was something which gave us the biggest boost , it was @nischaydnk great idea , so I will let him explain it in the comments.\n\n\nI take away a lot of learnings from this competition and I will be waiting eagerly for the next audio competition",
    "1332591": "Thanks @tanulsingh077 for sharing this!!\nAlso the reference in youtube is very helpful... the way I struggled in the competition even after looking at previous competition solutions made me realise I need to go back to the basics and have a good understanding of audio ML. This will be a useful reference! \ncongrats to you and the team on silver position!",
    "1335401": "Congratulation @tanulsingh077 for your 39th place and thanks for the detailed writing.\n\nThe idea of using deltas for stacking seems quite interesting. In terms of CV f1 points, how much did it helps? Also, did the two backbones models Resnext and Nfnet performed about the same, or one was better?",
    "1335956": "Thanks for sharing.\n\n> what really was more disappointing was that we failed to pick our best submission which was giving 0.66 on private\n\nHow would you have known this?  Have you learned something now that would have let you pick this submission?\n\nIt is interesting to see that 3 of the things that did not work work you did marvel for me.  Devil is in details probably.",
    "1336011": "> How would you have known this? Have you learned something now that would have let you pick this submission?\n\nNot before reading Psi's solution , our CV was same as everyone else's we used the train soundscapes 20 files to evaluate and chose the best threshold , but since we were training using 7 second clips and were also selecting the clips randomly , moreover our cv on train soundscapes correlated well with the lb , that made us careless and we didn't think much\n\n> It is interesting to see that 3 of the things that did not work work you did marvel for me. Devil is in details probably.\n\nYou are right devil is in the details , we might have made it work if only we had time , but I learned a lot and I am fortunate that I participated ,maybe I can do well in next year's competition",
    "1336013": "Hi @jonathanbesomi Using Delta our resnext cv improved from 0.712 --> 0.722 and our nfnet's cv increased from 0.728 ---> 0.730 \n\nNfnet was better than resnext , but their ensemble did not work for some reason and it gave bad cv as well as lb",
    "1336014": "Thanks @kamal basics are always the key to win kaggle competition",
    "1336432": "Hi @tanulsingh077, thanks for your reply. Wow, 0.01 improvement is really a lot! I've been trying training different models (similar to kkiller notebook) but had hard time to beat the 0.7 CV score. Would you mind make your code public? I would like to train one of you models to understand where I was missing out. Also, I'm pretty sure other Kagglers will find this useful as well. Thanks!"
  },
  "source": "meta"
}